Saved in:
Bibliographic Details
Main Authors: Yin, Xuanhua, Xu, Chuanzhi, Zhou, Haoxian, Wei, Boyu, Cai, Weidong
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.12575
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910051557441536
author Yin, Xuanhua
Xu, Chuanzhi
Zhou, Haoxian
Wei, Boyu
Cai, Weidong
author_facet Yin, Xuanhua
Xu, Chuanzhi
Zhou, Haoxian
Wei, Boyu
Cai, Weidong
contents Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.
format Preprint
id arxiv_https___arxiv_org_abs_2603_12575
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation
Yin, Xuanhua
Xu, Chuanzhi
Zhou, Haoxian
Wei, Boyu
Cai, Weidong
Computer Vision and Pattern Recognition
Diffusion Transformers (DiTs) are a dominant backbone for high-fidelity text-to-image generation due to strong scalability and alignment at high resolutions. However, quadratic self-attention over dense spatial tokens leads to high inference latency and limits deployment. We observe that denoising is spatially non-uniform with respect to aesthetic descriptors in the prompt. Regions associated with aesthetic tokens receive concentrated cross-attention and show larger temporal variation, while low-affinity regions evolve smoothly with redundant computation. Based on this insight, we propose AccelAes, a training-free framework that accelerates DiTs through aesthetics-aware spatio-temporal reduction while improving perceptual aesthetics. AccelAes builds AesMask, a one-shot aesthetic focus mask derived from prompt semantics and cross-attention signals. When localized computation is feasible, SkipSparse reallocates computation and guidance to masked regions. We further reduce temporal redundancy using a lightweight step-level prediction cache that periodically replaces full Transformer evaluations. Experiments on representative DiT families show consistent acceleration and improved aesthetics-oriented quality. On Lumina-Next, AccelAes achieves a 2.11$\times$ speedup and improves ImageReward by +11.9% over the dense baseline. Code is available at https://github.com/xuanhuayin/AccelAes.
title AccelAes: Accelerating Diffusion Transformers for Training-Free Aesthetic-Enhanced Image Generation
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2603.12575